Inverse identification of pipeline steel mechanical state by means of coupled magnetic measurements and artificial neural networks

Abstract

Artificial neural networks are widely used to develop models able to predict properties of interest by learning and establishing relationships between inputs and outputs of a system. It is particularly relevant for the inversion process in the framework of Non-Destructive Testing (NDT) involving strongly non-linear and non-monotonic behavior and/or saturations. The case-study considered in this work is a magnetic material subjected to uniaxial mechanical stress, plastic strain and magnetic field. An Artificial Neural Networks (ANN) model is proposed to predict the corresponding remanent magnetization, coercive field and the maximum magnetization as target properties. A series of experimental data made of various magneto-mechanical measurements are used to train, evaluate and validate the ANN model. The proposed model suitably predicts the magnetic properties of a second specimen of the material in the same magnetic field, plastic strain and stress ranges as the first specimen. An inverse ANN is then proposed to evaluate the mechanical loading and the plastic strain from the magnetic signature. Unique and accurate solutions are found that proves the relevance of machine learning approach in such NDT application.